B-Rep Model Classification Based On Deep Learning And Its Feature Visualization Research

With the rapid development of 3D modeling technology,3D models are widely applied in the fields of furniture,game animation,machine manufacturing,etc.The number of 3D models has increased greatly,which correspondingly has created difficulties to management and production design.Therefore,in order to solve the requirement of retrieval and reuse of existing 3D models,3D model classification technology emerged.How to identify 3D models in the perfect expression of 3D model shapes is a hot spot of current research.The traditional 3D model classification method mainly relies on the engineer's artificial design features for classification,which is time-consuming and has low classification accuracy.Owing to high dimensionality and complexity of 3D data,the calculation cost is large and the feature extraction process is complicated when expressing 3D shapes directly.Different from traditional classification methods,deep learning trains machine to learn features automatically to achieve classification by simulating human brain information processing mechanism.In the process of large-scale 3D model classification,the subjective judgment is reduced and the efficiency is obvious.This paper is based on the extensive application of deep learning algorithms,especially the convolutional neural network in the fields of image and natural language,to consider the introduction of deep learning algorithms in 3D model classification.The main research work is as follows:(1)Extracting muti-angle views to characterize the 3D model.Since the 3D model itself has complexity,there are input restrictions in the commonly used classification algorithms,and the model data needs to be transformed.As an intuitive description of the 3D model,the view is easy to access and can directly input the deep learning model.The model view extracted from a single perspective contains less information.We set the 3D model motion trajectory,renders the shot,and obtain the view of the 3D model from multiple angles.Then we organize and tile multiple views to include more complete 3D model information.(2)Proposing a method for 3D model classification based on convolutional neural network.Firstly,we construct a network model using convolutional neural network,and training model.Then,we use the trained model to extract the hierarchical features of the 3D model and complete the classification.In addition,we use the features extracted by each network layer of the model as shape indexes,and combine with the K nearest neighbor algorithm to compare the classification effects of different network layers.(3)Visualizing the hierarchical features extracted by(2).Firstly,the features extracted directly from the convolutional layer and the pooling layer are visualized to show the process of filter learning features.Then construct a deconvolutional network visualization to map the important features extracted by each layer of the network to the pixel space.Reconstruct the 3D model using visual features.Determine the network's "likes" of features,adjust the network model structure,and influence the final classification results.(4)The experimental results show that the improved model based on feature visualization can effectively improve the classification accuracy of B-Rep model and solve the classification problem of the 3D model.